Skip to content

Releases: unslothai/unsloth

Release list

Large Performance Gains + Fixes

Large Performance Gains + Fixes Latest
Latest

Choose a tag to compare

View all tags
@danielhanchen danielhanchen released this 09 Sep 15:18
v0.1.808-beta
ab28823
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

This is a large performance and reliability + bug fix release for Unsloth

Highlights

  • 1.2-1.7x faster diffusion. AMD 20% perf boost vs ROCM via Vulkan
  • 2x faster updating, remove SAC + AV false positives for Windows
  • Blender MCP, detect Hermes, AMD gibberish fixed (reported to AMD)
  • Over 250+ bug fixes, 60% smaller binaries and performance improvements
  • Strix iGPU BIOS popup - 3x faster inference if more VRAM for iGPU
  • Updated PyTorch to 2.11 from 2.10 - 2.14 will be soon. MLX improvements
  • Default to gpt-6-astra for Codex logins and improve OpenAI API processing
  • MLX fixed when self healing / updating, causing slowdowns for inf + training
  • Fix AppImage being very laggy. Fixed RAG upload issues stuck at 28% and API issues
  • New Docker image published at https://hub.docker.com/r/unsloth/unsloth for Studio and notebooks

Performance Boosts

  • Diffusion is 1.2x to 1.7x faster for INT8 / FP8 pathways - all models accelerated.
  • 23% faster prompt processing and 8% faster generation on Strix Halo.
  • Gated-delta models now train up to 25% faster on Apple Silicon.
  • Quantized MLX KV caches use up to 74% less prompt memory than before.

AMD + Windows

  • Strix Halo and Strix Point now default to Vulkan for faster inference.
  • AMD iGPUs without ROCm now use Vulkan instead of CPU on Linux.
  • Windows llama.cpp binaries are now signed to reduce Smart App Control blocks.
  • Windows now clearly explains when code integrity blocks model loading.
  • Reinstalling Unsloth on Windows keeps your supported PyTorch version.

MLX + Apple Silicon

  • Fresh Mac installs keep MLX training and exports working.
  • Fine-tune with DoRA and more DPO loss types on Apple Silicon.
  • Batched MLX generation now streams and samples each chat independently.
  • More multimodal models can be fine-tuned using text-only datasets.

Studio + API

  • RAG and document uploads are faster, show clearer progress and no longer appear stuck.
  • Studio menus, Find and Settings now open more smoothly.
  • GPT-6 Astra is available for Codex logins with Low through Max reasoning controls.
  • Improved OpenAI-compatible API streaming, audio input and model loading.
  • AppImage builds are more reliable with a pinned release toolchain.

Installs + Docker

  • PyTorch 2.11 is now the default across supported installers.
  • New NVIDIA Docker images for training and Studio, from Turing to Blackwell.
  • Native AMD64 and ARM64 images are available from Docker Hub.
  • The Unsloth Python package is now >60% smaller.

Docker: https://hub.docker.com/r/unsloth/unsloth

Download Unsloth Desktop

Unsloth Desktop is free and open source. Download it for:

  • Windows
  • macOS
  • Linux

🦥 Download Unsloth Desktop

What's Changed

Read more

Contributors

AronRubin, TimPietrusky, and 31 other contributors
Assets 8
Loading

Large Perf Improvements + Fixes

Large Perf Improvements + Fixes

Choose a tag to compare

View all tags
@danielhanchen danielhanchen released this 08 Sep 14:21
v0.1.807-beta
cedbb58
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

This is a large performance and reliability + bug fix release for Unsloth

Highlights

  • AMD uses Vulkan by default - 20% perf boost for prefill, decoding vs ROCM
  • Windows llama-server.exe is now signed, reducing false positives for SAC
  • AMD gibberish issues in Strix, iGPUs fixed in (upstream - reported to AMD)
  • Over 200+ bug fixes, 50% smaller binaries and performance improvements
  • Updated PyTorch to 2.11 from 2.10 - 2.14 will be soon. MLX improvements
  • Default to gpt-6-astra for Codex logins and improve OpenAI API processing
  • MLX fixed when self healing / updating, causing slowdowns for inf + training
  • Fix AppImage being very laggy. Fixed RAG upload issues stuck at 28% and API issues
  • New Docker image published at https://hub.docker.com/r/unsloth/unsloth for Studio and notebooks

AMD + Windows

  • Strix Halo and Strix Point now default to Vulkan for faster inference.
  • Up to 23% faster prompt processing and 8% faster generation on Strix Halo.
  • AMD iGPUs without ROCm now use Vulkan instead of CPU on Linux.
  • Windows llama.cpp binaries are now signed to reduce Smart App Control blocks.
  • Windows now clearly explains when code integrity blocks model loading.
  • Reinstalling Unsloth on Windows keeps your supported PyTorch version.

MLX + Apple Silicon

  • Gated-delta models now train up to 25% faster on Apple Silicon.
  • Quantized MLX KV caches use up to 74% less prompt memory than before.
  • Fresh Mac installs keep MLX training and exports working.
  • Fine-tune with DoRA and more DPO loss types on Apple Silicon.
  • Batched MLX generation now streams and samples each chat independently.
  • More multimodal models can be fine-tuned using text-only datasets.

Studio + API

  • RAG and document uploads are faster, show clearer progress and no longer appear stuck.
  • Studio menus, Find and Settings now open more smoothly.
  • GPT-6 Astra is available for Codex logins with Low through Max reasoning controls.
  • Improved OpenAI-compatible API streaming, audio input and model loading.
  • AppImage builds are more reliable with a pinned release toolchain.

Installs + Docker

  • PyTorch 2.11 is now the default across supported installers.
  • New NVIDIA Docker images for training and Studio, from Turing to Blackwell.
  • Native AMD64 and ARM64 images are available from Docker Hub.
  • The Unsloth Python package is now 44% smaller.

Docker: https://hub.docker.com/r/unsloth/unsloth

Download Unsloth Desktop

Unsloth Desktop is free and open source. Download it for:

  • Windows
  • macOS
  • Linux

🦥 Download Unsloth Desktop

What's Changed

Read more

Contributors

TimPietrusky, ianderse, and 26 other contributors
Loading

2x Faster Qwen3.8-Flash + GLM-5.3-Flash MTP

2x Faster Qwen3.8-Flash + GLM-5.3-Flash MTP

Choose a tag to compare

View all tags
@danielhanchen danielhanchen released this 02 Sep 12:49
v0.1.806-beta
d90c032
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

Run Qwen3.8-Flash-Next and GLM-5.3-Flash up to 2x faster with MTP. MTP is enabled by default, you can still disable it.
Also our new release includes 170+ training, chat, hardware, and performance improvements.

Highlights

  • Smoother model loading (less errors) across local servers and connected providers.
  • Faster and less laggy UI with follow-up turns much faster for all chats.
  • Safer chat edits that preserve tool cards, reply details, and conversation branches.
  • New local media APIs for video, audio, and MLX-served models.
  • New audio support with new models, progress tracking including: MiniMax-Music3, Higgs, MOSS and more!
  • Improved multi-GPU planning, memory fitting, and split-model training.
  • Strengthened AMD/ROCm detection, installation, and GPU compatibility.
  • Upgraded MCP, Deep Research, OAuth, and agent tool reliability.

Qwen3.8-Flash + GLM-5.3-Flash

  • Qwen and GLM now generate faster with MTP enabled by default.
  • Use GLM tools across longer, multi-turn chats.
  • Qwen automatically applies the recommended settings for thinking and non-thinking modes.

Download Qwen3.8-Flash-Next and GLM-5.3-Flash. See the Qwen guide and GLM guide for recommended settings and available GGUFs.

Faster MLX inference

  • Fine-tune both large MoE models with text or images on Apple Silicon using MLX.
  • Long Qwen chats now run much faster on Mac, with follow-up turns up to 30x faster.
  • MLX models now use their full context size and support much longer batched generation.
  • MLX releases GPU memory more cleanly between generation bursts and model switches.
  • Serve MLX models through Unsloth's OpenAI-compatible API.

Audio

  • Added support for MiniMax-Music3, Higgs, MOSS audio models.
  • Added live progress updates while audio is being generated.
  • Audio clips can now be archived and managed.
  • Improved reliability with custom TTS playback fixes, Whisper pairing checks, and stronger audio testing.

Chat + tools

  • Run several tool calls at once without mixing up their arguments.
  • Keep tools available when chatting with images.
  • Each chat keeps its MCP connection for faster tool calls.
  • Local models can edit code using Codex’s apply_patch tool.
  • Continue long chats with images and other media using Auto Compaction.
  • Review and approve Deep Research plans before research starts.

Training + hardware

  • Train larger models across multiple GPUs with automatic placement.
  • AMD installs choose the best build across Windows and Linux, with BF16 on more GPUs.
  • Export GLM-5.3 MLX fine-tunes to GGUF.
  • Choose custom GGUF shard sizes and save locations.

API + Desktop

  • Generate videos through the new OpenAI-compatible Videos API.
  • Updates download in the background and install when you restart.
  • Choose a custom port for LAN access.
  • Generate audio with Higgs, MOSS and MiniMax models.
  • Track audio generation progress and archive finished clips.
  • Model downloads show clearer progress and can switch from Xet to HTTP automatically.

Download Unsloth Desktop

Unsloth Desktop is free and open source. Download it for:

  • Windows
  • macOS
  • Linux

🦥 Download Unsloth Desktop

What's Changed

  • Put the smart offload planner back behind its flag by @danielhanchen in #9862
  • Studio: stop the per-chunk autosave writing back messages the server owns by @danielhanchen in #9865
  • Bump install.sh / install.ps1 pins to unsloth>=2026.8.22 by @danielhanchen in #9868
  • Fix Studio hydrating synced GGUF files before selection by @milewski in #9539
  • Fix datasets PyArrow registration after failed warm import by @Imagineer99 in #9626
  • fix(studio): stop offering a transformers upgrade where it cannot load anything by @Lyxot in #9878
  • Fix Studio CPT overwriting LFM2 all-linear LoRA targets by @Souravrajvi0 in #9870
  • fix(studio): accept trailing slash for model discovery by @Imagineer99 in #9876
  • studio: accept shared youtube clipboard links by @mahiatlinux in #9903
  • studio: tell users when a canvas is blocked by the network access setting by @NilayYadav in #9894
  • studio: fix streamed tool calls in studio api monitor by @Etherll in #9896
  • studio: serve installed mlx models through the openai-compatible api by @mahiatlinux in #8768
  • Fix sidebar footer width and context menu radius by @shimmyshimmer in #9918
  • strip inherited claude provider routing from local agent launches by @Souravrajvi0 in #9871
  • studio: honor a cancelled load inside the llama-server health wait by @NilayYadav in #9893
  • fix model hub recent sorting for on-device models by @Imagineer99 in #9642
  • add a setting that tells the model the current date by @mahiatlinux in #8879
  • studio: fix lan address actions and launch bind reporting by @Lyxot in #9170
  • rag: avoid re-hashing linked-folder snapshots in start_ingestion by @DhruvaMyakeri in #9913
  • fix(studio): coalesce concurrent keyless-settings reads to stop SQLite thundering herd by @Greninja44 in #9911
  • Studio: add the OpenAI videos API on /v1/videos by @NilayYadav in #9891
  • Clarify tool approval copy by @Imagineer99 in #8738
  • Pin Claude Code to the loaded model's real context window by @NilayYadav in #9940
  • studio: drop the schema bounds llama.cpp's grammar engine cannot compile by @Lyxot in #9923
  • Fix shell selection for remote Agents commands by @Imagineer99 in #8844
  • Studio: accept the loadConfig keys a saved system-prompt preset already sends by @indrajeetapache in #9881
  • Studio: compact long chats with media attachments by @Imagineer99 in #9476
  • Studio: add Higgs, MOSS, and MiniMax audio models by @Etherll in #8794
  • studio: add ordered arguments to mcp server settings by @Etherll in #9943
  • studio: list route-aware models through the openai api by @NilayYadav in #9892
  • fix(studio): repair custom TTS endpoint test and playback by @Souravrajvi0 in #9905
  • Studio: render escaped inline math in lists by @Imagineer99 in #9796
  • deprecate unsupported custom prompt templates by @vineethsaivs in #8938
  • Restore UNSLOTH_RETURN_LOGITS after prediction_step instead of forcing it to 0 by @danielhanchen in #6239
  • Studio: update in the background and restart when ready by @NilayYadav in #9890
  • Serve every MiniMax H3 asset from unsloth rather than a community repack by @danielhanchen in #8832
  • fix(studio): gate external tool nudges and preserve retry context by @Biotrioo in #9125
  • Chat: let the LoRA compare rehydrate a pair the generalized compare saved by @lonexreb in #9908
  • fix: expose codex apply_patch for local models (#9114) by @Biotrioo in #9121
  • Fix duplicate GGUF rows in custom model folders by @Imagineer99 in #9975
  • Revert "fix(studio): gate external tool nudges and preserve retry context" by @danielhanchen in #9990
  • studio: isolate mcp browser test state by @mahiatlinux in #9951
  • Fix three CI regressions: code-span anchor, load-dispatch anchor, recordless pip install by @danielhanchen in #9959
  • studio: fix audio ci tests by @mahiatlinux in #9953
  • studio: fix openai route ci tests by @mahiatlinux in #9952
  • Skip the inline-code mask when every backtick is inside a fence by @danielhanchen in #9991
  • stop the planner withholding a split reserve from one card, and price physical cores by @danielhanchen in #9863
  • Parse the downloaded model mapper instead of exec'ing it by @danielhanchen in #9777
  • Recapture the control leg reference against the fixed loss normalisation by @danielhanchen in #9569
  • Stop os.dup2 alone being a reverse shell, and allowlist the zoo URL guard by @danielhanchen in #9994
  • Give the in-venv CLI tests a writable STUDIO_HOME by @danielhanch...
Read more

Contributors

milewski, indrajeetapache, and 27 other contributors
Loading

2x Faster Qwen3.8-Flash + GLM-5.3-Flash MTP

2x Faster Qwen3.8-Flash + GLM-5.3-Flash MTP

Choose a tag to compare

View all tags
@danielhanchen danielhanchen released this 02 Sep 11:41
v0.1.805-beta
87592c5

Run Qwen3.8-Flash-Next and GLM-5.3-Flash up to 2x faster with MTP. MTP is enabled by default, you can still disable it.
Also our new release includes 170+ training, chat, hardware, and performance improvements.

Highlights

  • Smoother model loading (less errors) across local servers and connected providers.
  • Safer chat edits that preserve tool cards, reply details, and conversation branches.
  • New local media APIs for video, audio, and MLX-served models.
  • New audio support with new models, progress tracking including: MiniMax-Music3, Higgs, MOSS and more!
  • Improved multi-GPU planning, memory fitting, and split-model training.
  • Ctrl+F Search now works.
  • Strengthened AMD/ROCm detection, installation, and GPU compatibility.
  • Upgraded MCP, Deep Research, OAuth, and agent tool reliability.

Qwen3.8-Flash + GLM-5.3-Flash

  • Qwen and GLM now generate faster with MTP enabled by default.
  • Use GLM tools across longer, multi-turn chats.
  • Qwen automatically applies the recommended settings for thinking and non-thinking modes.

Download Qwen3.8-Flash-Next and GLM-5.3-Flash. See the Qwen guide and GLM guide for recommended settings and available GGUFs.

Faster MLX inference

  • Fine-tune both large MoE models with text or images on Apple Silicon using MLX.
  • Long Qwen chats now run much faster on Mac, with follow-up turns up to 30x faster.
  • MLX models now use their full context size and support much longer batched generation.
  • MLX releases GPU memory more cleanly between generation bursts and model switches.
  • Serve MLX models through Unsloth's OpenAI-compatible API.

Audio

  • Added support for MiniMax-Music3, Higgs, MOSS audio models.
  • Added live progress updates while audio is being generated.
  • Audio clips can now be archived and managed.
  • Improved reliability with custom TTS playback fixes, Whisper pairing checks, and stronger audio testing.

Chat + tools

  • Run several tool calls at once without mixing up their arguments.
  • Keep tools available when chatting with images.
  • Each chat keeps its MCP connection for faster tool calls.
  • Local models can edit code using Codex’s apply_patch tool.
  • Continue long chats with images and other media using Auto Compaction.
  • Review and approve Deep Research plans before research starts.

Training + hardware

  • Train larger models across multiple GPUs with automatic placement.
  • AMD installs choose the best build across Windows and Linux, with BF16 on more GPUs.
  • Export GLM-5.3 MLX fine-tunes to GGUF.
  • Choose custom GGUF shard sizes and save locations.

API + Desktop

  • Generate videos through the new OpenAI-compatible Videos API.
  • Updates download in the background and install when you restart.
  • Choose a custom port for LAN access.
  • Generate audio with Higgs, MOSS and MiniMax models.
  • Track audio generation progress and archive finished clips.
  • Model downloads show clearer progress and can switch from Xet to HTTP automatically.

Download Unsloth Desktop

Unsloth Desktop is free and open source. Download it for:

  • Windows
  • macOS
  • Linux

🦥 Download Unsloth Desktop

What's Changed

Read more

Contributors

milewski, indrajeetapache, and 27 other contributors
Loading

Qwen3.8-Flash-Next + GLM-5.3-Flash

Qwen3.8-Flash-Next + GLM-5.3-Flash

Choose a tag to compare

View all tags
@danielhanchen danielhanchen released this 27 Aug 13:09
v0.1.804-beta
8c43aed

Qwen3.8-Flash-Next and GLM-5.3-Flash can now run locally in Unsloth!

  • Run Qwen3.8-Flash on 75GB RAM, GLM-5.3-Flash on 102GB RAM+VRAM
  • 5x Faster inference for RAM offloading
  • "Infinite" repeated compaction now works
  • 100+ chat, reliability and performance improvements

Qwen Guide: https://unsloth.ai/docs/models/qwen3.8-next
Qwen GGUFs: https://huggingface.co/unsloth/Qwen3.8-Flash-Next-GGUF
GLM Guide: https://unsloth.ai/docs/models/glm-5.3-flash
GLM GGUFs: https://huggingface.co/unsloth/GLM-5.3-Flash-GGUF

Highlights

  • Qwen3.8-Flash-Next on 75GB RAM
  • GLM-5.3-Flash on 102GB total memory
  • Smarter GPU + RAM offloading - run larger models with less setup
  • Chats recover after disconnects instead of losing the reply
  • See what fits before loading with clearer memory estimates

Qwen3.8-Flash-Next

Qwen3.8-Flash-Next is a new 125B multimodal reasoning model and an early preview of Qwen4's architecture.

  • The 1-bit Unsloth Dynamic GGUF runs on 75GB RAM or unified memory.
  • It's 79% smaller than BF16 while retaining 80% top-1 accuracy.
  • Chat with text and images using up to 262K context.
  • Switch between None, Low, Medium and Extra High reasoning.
  • Preserved Thinking keeps reasoning consistent across longer chats.

GLM-5.3-Flash

GLM-5.3-Flash is Z.ai's new 320B multimodal model, with only 18B parameters active at a time.

  • Run the 1-bit model on 102GB of combined RAM + VRAM.
  • Chat with text, images and long documents using up to 1M context.
  • Switch between Low, High and Max reasoning.
  • Stronger coding, agent and vision performance than GLM-5.2.
  • Recommended settings are applied automatically in Unsloth.

Chat + tools

  • Local chats resume after a disconnect instead of losing the reply.
  • Deep Research keeps going when a provider asks it to slow down.
  • Vision chats now handle multiple images properly.
  • Images returned by MCP tools appear directly in chat.
  • Export chats as JSONL for backups or use in other tools.
  • Adjust Auto Compaction for longer chats, or turn it off.
  • Collapse tool activity by default for cleaner agent chats.

Models + performance

  • Large GGUFs automatically split across GPU and system RAM.
  • See estimated memory usage before loading a model.
  • View VRAM usage directly from your downloaded models.
  • Model settings stay saved when switching chats.
  • Search and download embedding models directly from Hugging Face.
  • Text-to-speech models only load when you actually use them.

Desktop + reliability

  • Linux voice recording fixed.
  • NVIDIA + Wayland interface freezes fixed.
  • AMD model loading crashes fixed.
  • llama.cpp models now load from Windows profiles with non-English characters.
  • Non-English web links now work properly as chat sources.
  • Desktop download links always point to the latest stable release.

What's Changed

Read more

Contributors

Lwrless, kozistr, and 20 other contributors
Loading

Bug Fixes + Auto compaction + LAN Remote Access

Bug Fixes + Auto compaction + LAN Remote Access

Choose a tag to compare

View all tags
@danielhanchen danielhanchen released this 25 Aug 13:12
v0.1.803-beta
e08c149

Thanks for the support for Qwen3.8-27B and Unsloth Desktop! This is a bug fix release with 170+ PRs.

  • MLX fixed - Some MLX and Mac runtimes did not run correctly
  • LAN API keyless / password-less + Keyboard shortcuts
  • XET / HTTP download toggle - clearer download progress
  • AMD bug fixes + 170 bug, reliability & performance fixes

Features

  • Auto Compaction (Experimental) for longer chats beyond context limits
  • Remote & LAN Access (Preview) for easy network access without Cloudflare links
  • Faster Chat - Improved streaming performance, reduced UI lag, and smoother long conversations.
  • Support for custom llama.cpp builds. Toggles for Cache RAM, Mmap, Mlock, Checkpoints, Spe Decoding KV Cache, Vision On / Off
  • Unsloth Dynamic v3.0 is released. New Qwen3.8-27B Dynamic v3.0 GGUFs deliver >10% higher top-1 accuracy compared to everyone else. Works with Unsloth.

Auto compaction (Experimental)

You can now do long chats and surpass a model's context limit, with evicted turns
remaining searchable.

  • Long local chats can continue past the context limit by rolling older turns out of
    active context.
  • Auto compaction creates fresh context epochs instead of permanently trimming replies.
  • Evicted conversations remain searchable, with improved fact recall.
  • Context window information is shown before chat begins.

How it works

  • Entire oldest turns are removed only when needed, never mid-message. The saved transcript stays unchanged.
  • Evicted turns are indexed into a per-thread searchable archive using Unsloth's existing RAG pipeline (store, chunker, embedder, retrieval). Lexical search is prioritized because chat recall is usually exact matches like names, numbers, or IDs.
  • A recall is forced during eviction instead of relying on the model to search itself. Later retrieval uses search_conversation.
  • The archive persists across epochs, allowing future compactions to recover previously evicted context.
  • No summarization: it showed little benefit and added ~190s per compaction.

Remote & LAN access (Preview)

Access Unsloth from another device on your network, managed from settings.

  • New dedicated Settings section for remote access.
  • Enable or disable LAN access without restarting.
  • Supports connection addresses, QR codes, and optional auto-start.
  • LAN access is disabled by default and requires changing the generated admin password.

Chat + Functionality improvements

  • Long chats are faster with optimized streaming and thread handling.
  • Projects organize chats, files, and workspaces together.
  • Chats remember composer settings, system prompts, and model sampling options.
  • Added prompt queueing and customizable keyboard shortcuts.
  • Added edit_file tool for partial file modifications.
  • Improved tool calling and MCP behavior.

Hardware + inference improvements

  • Support for custom llama.cpp builds.
  • Toggles for advanced settings - Cache RAM, Mmap, Mlock, Checkpoints, Speculative Decoding KV Cache, Vision On / Off
  • Better validation for missing GPU support and oversized GGUF models.
  • Improved MTP performance and VRAM handling.
  • Better compatibility across ROCm, xFormers, and flash-attention setups.
  • Added Intel XPU support.

API + serving

  • Responses API supports structured output formatting.
  • Improved recovery when llama-server crashes.
  • Added OpenCode V2 support for unsloth start.

Hub + Desktop + training

  • Hub supports more image/video workflows.
  • Faster and improved GGUF downloads.
  • Desktop packaging and permissions improved.
  • Training scheduler and preprocessing improvements.

What's Changed

Read more

Contributors

xyzs996, indrajeetapache, and 25 other contributors
Loading

Bug Fixes + Auto compaction + LAN Remote Access

Bug Fixes + Auto compaction + LAN Remote Access

Choose a tag to compare

View all tags
@danielhanchen danielhanchen released this 25 Aug 11:27
v0.1.802-beta
7e870b8

Thanks for the support for Qwen3.8-27B and Unsloth Desktop! This is a bug fix release with 170+ PRs.

  • MLX fixed - Some MLX and Mac runtimes did not run correctly
  • LAN API keyless / password-less is now supported
  • XET / HTTP download toggle - clearer download progress
  • AMD bug fixes for Strix Halo, all RDNA GPUs + 170 bug fixes

Features

  • Auto Compaction (Experimental) for longer chats beyond context limits
  • Remote & LAN Access (Preview) for easy network access without Cloudflare links
  • Faster Chat - Improved streaming performance, reduced UI lag, and smoother long conversations.
  • Support for custom llama.cpp builds. Toggles for Cache RAM, Mmap, Mlock, Checkpoints, Spe Decoding KV Cache, Vision On / Off
  • Unsloth Dynamic v3.0 is released. New Qwen3.8-27B Dynamic v3.0 GGUFs deliver >10% higher top-1 accuracy compared to everyone else. Works with Unsloth.

Auto compaction (Experimental)

You can now do long chats and surpass a model's context limit, with evicted turns
remaining searchable.

  • Long local chats can continue past the context limit by rolling older turns out of
    active context.
  • Auto compaction creates fresh context epochs instead of permanently trimming replies.
  • Evicted conversations remain searchable, with improved fact recall.
  • Context window information is shown before chat begins.

How it works

  • Entire oldest turns are removed only when needed, never mid-message. The saved transcript stays unchanged.
  • Evicted turns are indexed into a per-thread searchable archive using Unsloth's existing RAG pipeline (store, chunker, embedder, retrieval). Lexical search is prioritized because chat recall is usually exact matches like names, numbers, or IDs.
  • A recall is forced during eviction instead of relying on the model to search itself. Later retrieval uses search_conversation.
  • The archive persists across epochs, allowing future compactions to recover previously evicted context.
  • No summarization: it showed little benefit and added ~190s per compaction.

Remote & LAN access (Preview)

Access Unsloth from another device on your network, managed from settings.

  • New dedicated Settings section for remote access.
  • Enable or disable LAN access without restarting.
  • Supports connection addresses, QR codes, and optional auto-start.
  • LAN access is disabled by default and requires changing the generated admin password.

Chat + Functionality improvements

  • Long chats are faster with optimized streaming and thread handling.
  • Projects organize chats, files, and workspaces together.
  • Chats remember composer settings, system prompts, and model sampling options.
  • Added prompt queueing and customizable keyboard shortcuts.
  • Added edit_file tool for partial file modifications.
  • Improved tool calling and MCP behavior.

Hardware + inference improvements

  • Support for custom llama.cpp builds.
  • Toggles for advanced settings - Cache RAM, Mmap, Mlock, Checkpoints, Speculative Decoding KV Cache, Vision On / Off
  • Better validation for missing GPU support and oversized GGUF models.
  • Improved MTP performance and VRAM handling.
  • Better compatibility across ROCm, xFormers, and flash-attention setups.
  • Added Intel XPU support.

API + serving

  • Responses API supports structured output formatting.
  • Improved recovery when llama-server crashes.
  • Added OpenCode V2 support for unsloth start.

Hub + Desktop + training

  • Hub supports more image/video workflows.
  • Faster and improved GGUF downloads.
  • Desktop packaging and permissions improved.
  • Training scheduler and preprocessing improvements.

What's Changed

  • Bump install.sh / install.ps1 pin to unsloth>=2026.8.19 by @danielhanchen in #9422
  • Say what killed a UI wrapper script instead of reporting a bare 143 by @danielhanchen in #9407
  • Fix stale libappindicator3-dev guard blocking desktop releases by @danielhanchen in #9426
  • studio: drag and drop files onto project sources and knowledge bases by @mahiatlinux in #9147
  • Read a llama install's ggml tree from its release when the marker lacks one by @danielhanchen in #7912
  • studio: add custom stt connections by @mahiatlinux in #9349
  • Repair the two studio test suites #9410 left behind by @danielhanchen in #9437
  • CI: hold the Anthropic SDK below 1.0 in the inference smoke workflows by @danielhanchen in #9432
  • studio: collapse the reasoning pane with grid rows instead of a measured height, behind a flag by @danielhanchen in #9423
  • Studio: turn on the sidebar drag's scoped custom-property writes by @danielhanchen in #9441
  • Studio: re-ask the projector placement after a tensor-parallel downgrade by @danielhanchen in #9408
  • Studio: hide the Tensor Parallelism switch on diffusion models by @danielhanchen in #9418
  • Exclude the imatrix from the local models route too by @danielhanchen in #9411
  • Studio: fall back to the legacy --swa-checkpoints spelling by @danielhanchen in #9416
  • Retry the desktop ownership id read that a concurrent publish denies by @danielhanchen in #9459
  • Bound the SDK majors CI installs, and guard the ones the probes call by @danielhanchen in #9460
  • Stop the chart metric formatters truncating values of 1000 and up by @vineethsaivs in #8724
  • Studio: queue media generation behind model teardown by @divagr18 in #8866
  • Studio: keep server chats visible when legacy IndexedDB stalls by @oobabooga in #9446
  • fix(studio): rewind the MLX PRNG in the KV quantization probe without item assignment by @Lyxot in #9478
  • install: fetch managed Python when distro uv.toml sets downloads=manual by @gokay-ai in #9358
  • Studio: admit GGUFs that fit carved-out Vulkan UMA by @alkinun in #9461
  • Studio: settings nav badges and the Shortcuts icon by @shimmyshimmer in #9497
  • studio: stop a menu dismissal deleting a chat message by @mahiatlinux in #9243
  • Unwedge Backend CI: one deadlocked test, and a timer that names the next one by @danielhanchen in #9515
  • Studio: copy a thread selection without building the styled clipboard flavour by @danielhanchen in #9457
  • Studio: turn on the grid-based reasoning collapse by @danielhanchen in #9511
  • Studio: defer off-screen code fence highlighting, one way only by @danielhanchen in #9462
  • studio: measure the code block collapse the thread override prevents by @danielhanchen in #9464
  • Studio: accept .ods files from drag and drop by @Etherll in #9455
  • open every attachment type in a preview modal by @mahiatlinux in #8655
  • Import the OpenDocument MIME constants from where they are exported by @danielhanchen in #9520
  • Studio: don't let MCP training kill a live chat stream by @NilayYadav in #9434
  • Desktop: attach dropped source files on desktop instead of refusing them by @NilayYadav in #9495
  • Studio: avoid AppImage COLRv1 font crashes by @wasimysaid in #9473
  • Studio: stop an unterminated code fence re-repairing itself on every chunk by @danielhanchen in #9517
  • Studio: list the other embedding models, and show Documents & RAG in Data by @shimmyshimmer in #9514
  • Studio: document reasoning on the Agents tab, and add --reasoning-effort by @shimmyshimmer in #9513
  • Bound the backend pytest session so a wedged xdist worker cannot eat the job by @danielhanchen in #9530
  • Studio: paint a long thread's newest messages immediately instead of rebuilding all of them first by @danielhanchen in #9058
  • Allowlist the remote-server locale keys in the desktop branding sweep by @danielhanchen in #9529
  • Give the Studio T4 leg a Kaggle client that can read its own credential by @danielhanchen in #9535
  • Assert the event-loop property directly in the two load_freeze latency tests by @danielhanchen in #9532
  • studio: stream event feeds over post by @mahiatlinux in #9533
  • Studio: stop the KV starvation message reading as a context limit on the client by @danielhanchen in #9417
  • Studio: measure a dense tool result instead of estimating it, and never raise a configured cap by @danielhanchen in #9...
Read more

Contributors

xyzs996, indrajeetapache, and 25 other contributors
Loading

Auto compaction (preview) + LAN Remote Access

Auto compaction (preview) + LAN Remote Access

Choose a tag to compare

View all tags
@shimmyshimmer shimmyshimmer released this 20 Aug 16:58
v0.1.801-beta
f3a66cd

Thanks for the support for Qwen3.8-27B and Unsloth Desktop last week! For this release, we merged 200+ PRs to introduce many new features, fixes including:

  • Auto Compaction (Experimental) for longer chats beyond context limits
  • Remote & LAN Access (Preview) for easy network access without Cloudflare links
  • Faster Chat - Improved streaming performance, reduced UI lag, and smoother long conversations.
  • Support for custom llama.cpp builds. Toggles for Cache RAM, Mmap, Mlock, Checkpoints, Spe Decoding KV Cache, Vision On / Off
  • Unsloth Dynamic v3.0 is released. New Qwen3.8-27B Dynamic v3.0 GGUFs deliver >10% higher top-1 accuracy compared to everyone else. Works with Unsloth.

Auto compaction (Experimental)

You can now do long chats and surpass a model's context limit, with evicted turns
remaining searchable.

  • Long local chats can continue past the context limit by rolling older turns out of
    active context.
  • Auto compaction creates fresh context epochs instead of permanently trimming replies.
  • Evicted conversations remain searchable, with improved fact recall.
  • Context window information is shown before chat begins.

How it works

  • Entire oldest turns are removed only when needed, never mid-message. The saved transcript stays unchanged.
  • Evicted turns are indexed into a per-thread searchable archive using Unsloth's existing RAG pipeline (store, chunker, embedder, retrieval). Lexical search is prioritized because chat recall is usually exact matches like names, numbers, or IDs.
  • A recall is forced during eviction instead of relying on the model to search itself. Later retrieval uses search_conversation.
  • The archive persists across epochs, allowing future compactions to recover previously evicted context.
  • No summarization: it showed little benefit and added ~190s per compaction.

Remote & LAN access (Preview)

Access Unsloth from another device on your network, managed from settings.

  • New dedicated Settings section for remote access.
  • Enable or disable LAN access without restarting.
  • Supports connection addresses, QR codes, and optional auto-start.
  • LAN access is disabled by default and requires changing the generated admin password.

Chat + Functionality improvements

  • Long chats are faster with optimized streaming and thread handling.
  • Projects organize chats, files, and workspaces together.
  • Chats remember composer settings, system prompts, and model sampling options.
  • Added prompt queueing and customizable keyboard shortcuts.
  • Added edit_file tool for partial file modifications.
  • Improved tool calling and MCP behavior.

Hardware + inference improvements

  • Support for custom llama.cpp builds.
  • Toggles for advanced settings - Cache RAM, Mmap, Mlock, Checkpoints, Speculative Decoding KV Cache, Vision On / Off
  • Better validation for missing GPU support and oversized GGUF models.
  • Improved MTP performance and VRAM handling.
  • Better compatibility across ROCm, xFormers, and flash-attention setups.
  • Added Intel XPU support.

API + serving

  • Responses API supports structured output formatting.
  • Improved recovery when llama-server crashes.
  • Added OpenCode V2 support for unsloth start.

Hub + Desktop + training

  • Hub supports more image/video workflows.
  • Faster and improved GGUF downloads.
  • Desktop packaging and permissions improved.
  • Training scheduler and preprocessing improvements.
Platform Link
Windows Download
macOS Download
Linux / Ubuntu (deb) Download
Linux (AppImage) Download
Linux (Arm64) Download

What's Changed

  • Bump install.sh / install.ps1 pin to unsloth>=2026.8.18 by @danielhanchen in #8836
  • Give the llama-flag fixtures the parallelSlotsClamped they now require by @vineethsaivs in #8837
  • Studio: preserve pre-tool reasoning inside the GGUF tool loop by @losuler in #8581
  • docs: fix dead Linux .deb download link by @yzxcj797 in #8891
  • Model hub: show the Meta mark on the Unsloth re-upload of Muse Glimmer by @shimmyshimmer in #8949
  • Studio: speed up the local model inventory and keep it off the API loop by @oobabooga in #8770
  • Studio: reuse the cached GGUF work a load request has already done by @oobabooga in #8771
  • studio: keep each tool-call round separate when the delta index restarts by @mahiatlinux in #8754
  • studio: route bare tool-call fragments to the call that owns the index now by @mahiatlinux in #8755
  • studio: opt-in model auto-switch for the image and video apis by @mahiatlinux in #8766
  • Studio: keep project chats in Recents, add a Projects sidebar section by @shimmyshimmer in #8932
  • Studio: find the real Documents folder on Windows for project workspaces by @shimmyshimmer in #8955
  • fix(unsloth studio): only preprocess the rows a max_steps run will actually use instead of preprocessing validated datasets by @LeoBorcherding in #8890
  • Studio: stop the chat UI falling behind a fast stream by @oobabooga in #8845
  • studio: keep each chat's composer pills and settings with the chat by @mahiatlinux in #8686
  • studio: allow a per-connection max tokens limit on every provider by @mahiatlinux in #8944
  • Crash guard: read dumpability in order, and stop losing crashes to scope and naming by @danielhanchen in #8808
  • Studio: size the max_steps row bound for the data-parallel world by @danielhanchen in #8958
  • Studio: make the paste-to-attachment threshold a chat setting by @shimmyshimmer in #8963
  • Fix CI on main: stale test doubles, a stale router stub, and two source defects by @danielhanchen in #8956
  • Make the startup profile a gate, with budgets from its own measurements by @danielhanchen in #8965
  • Keep xFormers working when flash-attn 4 is installed, and guard the varlen int32 overflow by @danielhanchen in #8957
  • Stop six backend tests waiting on the clock instead of on a signal by @danielhanchen in #8975
  • CI: stop three workflows paying for work they throw away by @danielhanchen in #8976
  • Give the pasted-text import an extension so the node test runner can load it by @danielhanchen in #8981
  • Make the packed-boundary guard reachable on the fused cross-entropy path by @danielhanchen in #8959
  • Studio: Fix embedded MTP performance under partial GPU offload by @oobabooga in #8875
  • Studio: tokenize the dataset online for plain-text single-pass runs by @danielhanchen in #8960
  • Studio: keep K and V equal when the flash-attention retry resets the V cache by @danielhanchen in #8827
  • Studio: hold the Auto-layers context exemption only while --fit runs by @danielhanchen in #8825
  • Studio: stop a sent prompt staying in the chat composer by @shimmyshimmer in #8849
  • Chat: queue prompts with Cmd/Ctrl+Enter and drag to reorder the queue by @shimmyshimmer in #8952
  • Studio: repin the permission-mode contract to behaviour, not a declaration by @danielhanchen in #8986
  • Cut Repo tests (CPU) roughly in half: stop the synthetic vLLM tests reaping a fake server by @danielhanchen in #8993
  • Core: sweep every transformers model_type across a process pool by @danielhanchen in #8995
  • Repair the Tauri retry test against the download-marker pipeline by @danielhanchen in #8996
  • Studio: read every launcher world size when resolving a step-capped run's passes by @danielhanchen in #9000
  • Repair the prompt-queue contract test against the queueing refactor by @danielhanchen in #9026
  • studio: stop prom...
Read more

Contributors

indrajeetapache, losuler, and 21 other contributors
Loading
1 Join discussion

Qwen3.8-27B

Qwen3.8-27B

Choose a tag to compare

View all tags
@danielhanchen danielhanchen released this 14 Aug 14:18
v0.1.800-beta
a8be2a8
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

Qwen3.8-27B and Qwen3.8-2.4T can now be run locally in Unsloth!
Run on 17GB RAM via Unsloth Dynamic GGUFs. You can also fine-tune Qwen3.8-27B in Unsloth.
Qwen3.8-27B is by far the strongest model for its size. We also uploaded NVFP4 quants.

Guide: https://unsloth.ai/docs/models/qwen3.8
GGUF: https://huggingface.co/unsloth/Qwen3.8-27B-GGUF

See 1-bit Qwen3.8-2.4T GGUF running in Unsloth:
qwen3.8 unsloth desktop

Highlights

  • Qwen3.8-27B + extra llama-server arguments allowed + custom VRAM toggle
  • External provider has tool calling + tool support + login with Codex
  • Fast FP8 10x faster MiniMax-H3 inference (3 minutes vs 30)
  • 10% faster inference for GGUFs + Bypass permissions fixed

Chat + tools

  • Connected AI providers can use their own Search or Unsloth Desktop's built-in Search and tools. Tool results are passed back to the model so it can continue multi-step tasks.
  • Sign in with a Codex subscription and use Codex tools inside Chat.
  • Chat shows live prompt and generation speeds, while long streaming replies use much less CPU.
  • Chat settings stay with the conversation across remote sessions.
  • Paste a YouTube link to attach its transcript, including the title, channel, duration, link, and caption language.
  • Save a full chat or reply into your project's sources while keeping its reasoning, tool calls, and citations.

MiniMax-H3

  • MiniMax-H3 can run on smaller supported GPUs by splitting large model parts into pieces that fit.
  • The model picker now hides H3 options that the current hardware cannot run instead of letting them fail after selection.
  • H3 options are labelled Fast FP8 or Slow, making the large speed difference clear before downloading.

Performance + hardware

  • Inference is up to 10% faster in supported cases, with lower VRAM use and a tunable memory limit.
  • Idle image and video models can optionally unload to free VRAM for Chat or Training.
  • Added better support for AMD RDNA 3, RDNA 4, and Strix Halo systems. VRAM checks no longer reserve extra GPU memory.
  • Multi-GPU ROCm device matching is safer.
  • Macs now choose context size from the memory that is actually free.
  • RAG document indexing uses the CPU by default, so it no longer leaves a large GPU memory block reserved.
  • Fixed GGUF image detection when choosing a model for the API.

Custom llama.cpp arguments

  • Model settings now include an Extra Arguments box for custom llama-server flags.
  • Unsloth checks flags against the installed build and saves valid ones per model for normal, startup, and API loads. Flags that could break model loading or app security are rejected with a clear message.

Training + exports

  • Training checks whether a model needs a newer Transformers version and asks before installing it. The preview also tells you when the model must use 16-bit instead of 4-bit.
  • GGUF exports check disk space before starting a long merge, with clearer errors showing how much space is needed.
  • On Kaggle, large saves can use the bigger temporary drive when the working folder is too small.
  • Fixed LoRA being ignored by vLLM GRPO rollouts on newer TRL versions.

Debugging + reliability

  • Settings > Debugging now shows the main server and model-runner logs, with live, 3-second, and manual refresh modes.
  • Copy the visible log or its full path for a bug report. Passwords, tokens, and API keys are hidden in the viewer, while files on disk stay unchanged.
  • Cached RAG models load first, avoiding long stalls when Hugging Face is unreachable.
  • Slow installs can keep running during large downloads and show what they are waiting for.
  • Fixed Hugging Face dataset links and linked folders on Windows.
  • Pinned models can be reordered by dragging them in the model list.

Download Unsloth Desktop

Unsloth Desktop is free and open source. Download it for:

  • Windows
  • macOS
  • Linux

🦥 Download Unsloth Desktop

Star Unsloth on GitHub

What's Changed

Read more

Contributors

danielhanchen, Lyxot, and 10 other contributors
Loading

v0.1.702-beta

v0.1.702-beta

Choose a tag to compare

View all tags
@danielhanchen danielhanchen released this 13 Aug 15:21
v0.1.702-beta
662fead

Unsloth Desktop is here! The first desktop app to run and train AI models locally. Research, export and deploy from the same open-source app on Windows, macOS and Linux.

v0.1.702-beta Update (August 13th)

  • Added tool calling / web search & more for all external providers
  • Fixed bypass permissions not working for sandboxing
  • UI and UX fixes - VRAM usage is now tunable
  • 10% faster inference + reduced VRAM usage and other perf fixes
  • Much better AMD RDNA3,4 + Strix Halo, Mac support
  • Image diffusion, video generation fixes
  • Can login with Codex subscription
  • Many bug fixes

unsloth desktop

🦥 Download Unsloth Desktop for Linux, Windows, MacOS

Here's what you can do with Unsloth Desktop:

  • Get up to 50% more accurate tool calling with self-healing calls and sandboxed code execution.
  • Run Muse Glimmer 30B, Kimi K3, Qwen3.8, DeepSeek-V4 Flash 0731, Gemma 4, and more.
  • Generate videos with MiniMax-H3, and create images and videos with other diffusion models at up to 2× faster inference on supported workflows.
  • Use unlimited private web search, Deep Research, RAG and MCP.
  • Export models to NVFP4, GGUF and other formats.
  • Access Unsloth remotely through Cloudflare HTTPS.
  • Run on CPU or multiple GPUs across NVIDIA, AMD, Intel and Mac.
  • Train models without code, using less time and VRAM.
  • Use local models through Unsloth's OpenAI-compatible API, or connect OpenAI and Anthropic models.

Tools, private research + APIs

Self-healing tool calling repairs malformed calls instead of dropping them. Models can run Python and Bash inside sandboxed environments, so they can test code, create files and verify their work.

Use unlimited private web search, let Deep Research plan and produce cited reports, or bring your own files into RAG. You can also connect MCP tools for workflows that need external apps, data or actions.

Local models can be served through Unsloth's OpenAI-compatible API for agents and other clients. Inside Desktop, you can also connect OpenAI and Anthropic as cloud model providers.

Muse Glimmer 30B + latest models

Run Muse Glimmer 30B locally for chat, agents, tools and APIs, alongside Kimi K3, Qwen3.8, DeepSeek-V4 Flash 0731 and Gemma 4. Download and manage them in one place through Unsloth Desktop.

MiniMax-H3 + image and video diffusion

Run MiniMax-H3 locally for video generation. Create images and videos locally, edit existing images and train supported diffusion models. Use LoRAs, reference images and ControlNet where available, with up to 2× faster inference on supported workflows.

No-code training, export + remote deployment

Pick a model and dataset, adjust the settings and start training. You can train supported LLMs, diffusion models, TTS models and embedding models without writing code. On supported LLM workloads, training is up to 2× faster and uses up to 70% less VRAM.

Export your trained models to NVFP4, GGUF and other supported formats. You can also securely deploy and access models remotely: turn on Remote access to publish Unsloth through a Cloudflare HTTPS link, then use the app and its local APIs from another device.

Hardware + platform support

Unsloth Desktop runs on Windows, macOS and Linux. Hardware support spans CPU and multi-GPU systems, NVIDIA and AMD GPUs, Intel hardware, and Mac.

CPU support includes Chat and Data Recipes. Training and inference options vary by model and backend.

Download Unsloth Desktop

Unsloth Desktop is free and open source. Download it for:

  • Windows
  • macOS
  • Linux

🦥 Download Unsloth Desktop

Star Unsloth on GitHub

v0.1.701-beta Update (1PM PDT time August 11th)

  • Fixed slow Windows downloading (200x faster due to throttling)
  • Fixed Mac asking to download command line tools (uv bug)
  • Fixed AMD Strix Halo not being detected
  • Other bug fixes
Loading